Proc. of the EAA Joint Symposium on Auralization and Ambisonics, Berlin, Germany, 3-5 April 2014
PERCEPTUAL AND ROOM ACOUSTICAL EVALUATION OF A COMPUTATIONAL
EFFICIENT BINAURAL ROOM IMPULSE RESPONSE SIMULATION METHOD
Torben Wendt,
Acoustics Group, Medical Physics
and Cluster of Excellence “Hearing4all”,
University Oldenburg
Oldenburg, Germany
Steven van de Par,
Acoustics Group and Cluster of Excellence “Hearing4all”,
University Oldenburg
Oldenburg, Germany
Stephan D. Ewert,
Medical Physics and Cluster of Excellence “Hearing4all”,
University Oldenburg
Oldenburg, Germany
ABSTRACT
A fast and perceptively plausible method for synthesizing bin-
aural room impulse responses (BRIR) is presented. The method
is principally suited for application in dynamic and interactive
evaluation environments (e. g., for hearing aid development), psy-
chophysics with adaptively changing room reverberation, or simula-
tion and computer games. In order to achieve a low computational
cost, the proposed method is based on a hybrid approach. Using the
image source model (ISM; Allen and Berkley [J. Acoust. Soc. Am.
Vol. 66(4), 1979]), early reflections are computed in a geometrically
exact way, taking into account source and listener positions as well
as wall absorption and room geometry approximated by a “shoe-
box”. The ISM is restricted to a low order and the reverberant tail
is generated by a feedback delay network (FDN; Jot and Chaigne
[Proc. 90th AES Conv., 1991]), which offers the advantages of a
low computational complexity on the one hand and an explicit con-
trol of the frequency dependent decay characteristics on the other
hand. The FDN approach was extended, taking spatial room proper-
ties into account such as room dimensions and different absorption
characteristics of the walls. Moreover, the listener orientation and
position in the room is considered to achieve a realistic spatial
reverberant field.
Technical and subjective evaluations were performed by com-
paring measured and synthesized BRIRs for various rooms. Mostly,
a high accuracy both for some common room acoustical parameters
and subjective sound properties was found. In addition, an analysis
will be presented of several methods to include room geometry in
the FDN.
1. INTRODUCTION
Room acoustical simulations are desirable for many purposes, such
as developing or testing signal processing algorithms, or to e. g.
test the effect of reverberation on speech intelligibility. Further-
more, they are of interest for audio-visual simulation environments
(e. g, for training and rehabilitation) and in entertainment, e. g. in
computer games, all requiring a real-time adaptation of the virtual
environment, depending on the movement of the listener and/or the
sound sources.
One traditional way to emulate the acoustics of a certain room
is to measure binaural room impulse responses (BRIRs) and to
convolve dry source signals with the BRIRs. However, such mea-
surements are time consuming and their usage is restricted to static
scenarios. Furthermore, one is restricted to actually existing rooms.
Alternatively, room acoustics can be simulated, enabling different
degrees of realism, ranging from simple artificial reverb generation
to complex room acoustical simulation (image source model [
1
],
CATT [
2
], ODEON [
3
]), even for dynamic scenarios (e. g. [
4
], [
5
],
[6]).
Depending on the application, physically correct rendering
of a soundfield is required or a perceptually convincing auraliza-
tion, implying plausibility and authenticity, is sufficient. For room
simulations used in psychoacoustic research, rehabilitation or in
computer games perceptual aspects are most important, implying
accordance of room acoustical parameters, e. g. reverberation time,
definition, and measures like speech intelligibility. In this case sim-
plifications can be made to reach computational efficiency allowing
for real-time rendering of dynamic acoustic scenes, in which the
positions of sources and receivers can be changed interactively.
Several approaches exist to synthesize room impulse responses.
If the wavelength of a sound is small compared to the characteristic
dimensions of reflecting objects, concepts of geometric acoustics
(GA), such as the image source model [
1
] or the ray tracing method
[
7
] can be applied. Both methods have been used and further
developed in various room acoustics simulation algorithms, mostly
as hybrids together with other algorithms (e. g. [
3
], [
8
]). However,
these methods still have high computational complexities.
If the exact room geometry is neglected, artificial reverberation
can be synthesized very efficiently and with predefined reverbera-
tion time. Here, a common approach are feedback delay networks
(FDNs), based on Schroeder’s pioneering work on parallel delay
lines with feedback [9] and further developed (amongst others) by
Stautner and Puckette [10] and Jot and Chaigne [11].
One way to achieve real-time performance while maintaining
86
Proc. of the EAA Joint Symposium on Auralization and Ambisonics, Berlin, Germany, 3-5 April 2014
the advantages of the more “accurate” GA-based BRIR synthe-
sis and reverberation algorithms is their combination in a hybrid
approach: The initial part of the impulse responses is computed
based on a GA method. The reverberant tail is generated by a more
effective reverberation algorithm. Perception motivates such an
approach as the early sound reflections create the impression of a
certain spatial source width on the one hand and support speech
intelligibility on the other hand. The following reverberant tail
contains diffuse reflections and its frequency dependent decay char-
acteristics conveys information about the wall absorption and room
size.
Here, a hybrid approach was evaluated which combines the
image source model (ISM) for a shoebox geometry to simulate early
reflections up to a low order, and an FDN for creating a diffuse
reverberation tail. The FDN was extended to be directly linked
to the room geometry used in the ISM, and to be able to spatially
render the reverberation tail in order to generate BRIRs. For low
ISM orders, BRIRs can be simulated very efficiently with this
approach. In technical and subjective evaluations, the ability of the
algorithm to create plausible and authentic simulations was assessed
for single and connected (coupled) shoebox rooms. Two different
approaches for spatial reverb distribution rendering were compared,
taking room dimensions and receiver position into accout.
2. SIMULATION METHOD
A hybrid approach [
12
] was used to synthesize BRIRs. Early sound
reflections are computed by an image source model up to a low
order. The late reverberation was generated by a feedback delay
network.
The auralization steps are described explicitly for the case
of headphone presentation, reflected by the application of head-
related impulse responses (HRIRs). The adaptation to arbitrary
loudspeaker-based playback systems, such as higher order ambison-
ics or wave field synthesis, can be easily achieved by replacing
them by respective loudspeaker-controlling functions.
2.1. Image source model
The ISM regards a sound reflection as the direct sound of a mirrored
version of the original source. This so-called image source differs
from the original source by its time delay and its attenuation due to
the distance to the receiver, as well as the respective wall reflection
coefficient. The sound of an image source is reflected again at other
walls, creating higher order image sources. In this way, arbitrarily
complex reflection paths can be modeled.
The ISM implementation in the proposed simulation method
is restricted to empty shoebox-shaped rooms, where the six wall
surfaces are represented each by frequency dependent absorption
coefficients. These shoebox-shaped rooms enable a very efficient
calculation of image source positions in comparison to arbitrary
room geometries [
13
]. Nevertheless, for a shoebox room the num-
ber of image sources up to reflection order
N
is of order
O(N3)
,
which considerably affects computational efficiency for higher re-
flection orders. Another limitation of the ISM is that it inherently
assumes only specular instead of diffuse reflections, although they
are of importance to describe room acoustics.
In the ISM implementation, the following signal processing
steps are performed for each image source: A “1
/
distance” at-
tenuation factor and a time delay due to distance to receiver; an
“effective reflection filter”, being the (frequency domain) product
of all wall reflection coefficients that are involved to “create” the
current image source; an HRIR, according to the azimuth and ele-
vation position of the image source relative to the receiver’s head
orientation. Finally, the binaural signals for all image sources are
added up to one two-channel output.
2.2. Extended feedback delay network
The extended FDN used here is based on the general multichannel
network as suggested by Jot and Chaigne [
11
] and consists basically
of a set of parallel delay lines whose outputs are fed back via a
feedback matrix A.
The number of parallel channels (delay lines) was set to 12,
with four channels associated to each (showbox) room dimension
(two channels per wall) reflected in several parameter choices.
Firstly, the delay units
τj, j ∈N≤12
were directly related to the
room dimensions via sound propagation speed (plus a random jitter
per channel). Secondly, the absorption filters with transfer functions
Habs
j
simulate the frequency dependent sound attenuation due to the
wall reflections and air absorption. After Jot and Chaigne [
11
] the
frequency dependent reverberation time
T60(f)
conveyed by the
resulting RIR is controlled explicitly by the following frequency
responses, if all other processing steps are energy preserving:
20 lg |Habs
j(f)|=−60τj/T60(f).(1)
In the simulation method, the reverberation time is predicted from
the wall absorption coefficients via Sabine’s formula. Thirdly, the
feedback matrix
A
redistributes the outputs back to the input chan-
nels. This process is energy preserving if
A
is an orthogonal matrix.
Here, a randomly created unitary matrix was chosen, providing a
high variety of pulse amplitudes.
Two last processing steps per channel, referred to as “binaural-
ization steps”, extend the FDN to introduce spatiality distributed
and externalized reverberation. (1) Via HRIR filtering the FDN
channels are mapped to 12 points (directions) around the head,
with two points positioned on each wall. (2) Reflection filters—
identical to those applied to the first order image sources in the
ISM—simulate a direction dependent sound intensity of reverbera-
tion, due to the different acoustical wall properties.
Two possible principles are suggested to map the 12 directions
around the receiver’s head, which are sketched in Fig. 1 (micro-
phone symbol: recveiver, big “
⊗
”: direct sound source, small “
⊗
”:
reverb source). The first one (lhs of Fig. 1) is called “cube” condi-
tion. Here, the 12 directions are mapped to points on a cube around
the receiver’s head. The cube always moves with the receiver (re-
ceiver is always in its centre) and is axis aligned with the room.
Figure 1: Illustration of two possible techniques of spatial reverb
distribution: “cube” (lhs) and “box” (rhs) method. See text for
explanation. (Arrows will be explained in sec. 3.2.
87
Proc. of the EAA Joint Symposium on Auralization and Ambisonics, Berlin, Germany, 3-5 April 2014
In this way all 12 incidence directions are more or less equally
distributed around the head. In the second “box” condition the 12
directions are mapped to points on the actual six wall surfaces like
depicted in the rhs part of Fig. 1. Here, the sound incidences are
warped according to the room dimensions and the actual receiver
position. Differences between both methods will clearly be audible
for rooms with large differences in dimensions. If not specified, in
the following the cube condition will be used as standard rendering
method.
2.3. Combination of ISM and FDN
For a smooth transition from the early-reflections part (ISM) to the
late-reverberation part (FDN), i. e. a straight decay of the BRIR
on dB scale, the energy and initial delay of the FDN input signal
have to be suitable. For this purpose the FDN input signal consists
of the
N
th-order ISM pulses before HRIR filtering. In order to
avoid comb-filter coloration effects, which occur if a fixed temporal
pattern of pulses is fed into the FDN, the ISM output is distributed
to the FDN channels. Because the number of image sources of
order
N
does in general not equal the number
M
of FDN channels,
the
i
th ISM pulse is fed into the FDN channel
[(i−1)modM] + 1
.
2.4. Simplified model for coupled rooms
In addition to the single shoebox-shaped room a strongly simplified
method to simulate the acoustics of two connected shoebox rooms
that are acoustically coupled, e. g. by an open door is suggested.
It is assumed that a source S is located in room 1 and a re-
ceiver R in room 2 as depicted in Fig. 2. The sound transmission
from room 1 to room 2 is then simulated by a single virtual source S
0
located in the door which is exciting room 2 as in the case of the
single shoebox simulation described above. The virtual source radi-
ates the monaural impulse response of room 1 for a source position
specified by the coupled-room arrangement and a “monaural” re-
ceiver R
0
inside the open door. Thus, the effective BRIR is obtained
as the convolution product of the monaural RIR of room 1 with the
BRIR of room 2.
Depending on the source position in room 1, it is either visible
or invisible for a receiver in room 2. If it is not visible, no direct
sound will arrive at the receiver but only reflections and diffractions.
In this case, the direct sound pulse of the RIR of room 1 is discarded
in the current approach.
Figure 2: Sketch of two adjacent rooms, that are acousticly coupled
by an open door. See text for explanation. (The arrow will be
explained in sec. 3.2.)
3. EVALUATIONS
Three main aspects of the proposed simulation method were eval-
uated. Firstly, for a set of real-existing rooms, subjectively rated
sound properties of measured and respectively synthesized BRIRs
were compared. Secondly, the two approaches to realize the binau-
ralization steps of the extended FDN (see sec. 2.2) were evaluated
with respect to binaural parameters. Thirdly, the approach to simu-
late the acoustics of two coupled rooms was evaluated.
To perform these evaluations, a test-database containing mea-
sured and synthesized BRIRs was created. BRIRs were measured
for various rooms of different size and reverberation time, as well
as a for a few source-receiver configurations in two connected
rooms. Additionally, some measured BRIRs were taken from the
AIR database [14].
The BRIR measurements were performed using an omnidirec-
tional loudspeaker based on a ring-radiator and an artificial head
MK2 by Cortex. Rooms were excited with a logarithmic sweep
[
15
] (50 Hz to 18 kHz) offering removal of nonlinear harmonic dis-
tortions from the recorded and inverse filtered signal. BRIRs were
each calculated as the mean of BRIRs from 10 single recordings
and equalized by the inverse loudspeaker transfer function.
For the BRIR synthesis a single mean wall absorption coeffi-
cient was used. It was determined for each room from its reverbera-
tion time via the inverse form of Sabine’s formula, which ensures
that reverberation times of measured and synthesized BRIRs are in
good accordance.
The HRIRs used in the simulation were from the same arti-
ficial head as used for the BRIR recordings. The database [
16
]
offers HRIRs with azimuth angles in 2
◦
steps at elevations near the
equatorial level. Towards the poles, the azimuth angle sampling
decreases. Elevation angles are sampled in 2◦steps.
A varying synthesis parameter was the maximum image source
order. The goal was to find a trade-off between accuracy and
computational efficiency.
In the following, for all different rooms and source-receiver
configurations, the term “room condition” will be used. In contrast,
different types of BRIR synthesis, differing in the choice of simula-
tion parameters, will be referred to as “synthesis conditions”. All
room- and synthesis conditions will be introduced in the following.
3.1. Subjective sound properties
3.1.1. Room- and synthesis conditions
BRIRs were chosen from four rooms of different size and rever-
beration time, specified in Tab. 1. For the synthesized BRIRs, the
maximum image source order
N
was varied in
{1,3}
. For one
room the BRIR was synthesized only by the ISM with
N= 20
. In
the extended FDN, the cube condition was chosen (see 2.2). Two
dry source signals, female spoken speech and a guitar play (steel
strings) were convolved with the recorded and simulated BRIRs.
Presentation sound pressure levels ranged from 60 to 65 dB SPL,
depending on the source-receiver distance and the room reverbera-
tion.
3.1.2. Subjects and procedure
15 normal-hearing subjects (7 female, 8 male) aged 24 to 32 years
participated in the experiment. Sounds were presented via head-
phones in a sound attenuating booth. Since the synthesis method
was implemented as an offline simulation, no head tracking and
adaptively changing soundfield was employed.
The sound properties which were to be rated on a seven-point
scale were “naturalness” and “room size”. A test and a retest were
88
Proc. of the EAA Joint Symposium on Auralization and Ambisonics, Berlin, Germany, 3-5 April 2014
Table 1: Rooms, whose BRIRs were used in the subjective evalua-
tion. Reverberation times T60 were obtained from measured BRIR
(broadband).
Room Dimensions (m) T60 (s)
Aula (12.0, 30.0, 10.0) 4.8
Empty chamber (1.88, 2.74, 2.82) 2.5
Lecture room (10.90, 10.80, 3.15) 0.8
Laboratory (4.97, 4.12, 3.00) 0.3
performed in two sessions, each with a randomized order of pre-
sented sounds. Before the actual experiment was performed, sound
examples illustrating extremal distinctions of the sound properties
had been presented.
3.1.3. Results
Fig. 3 shows the results from the subjective sound property ratings
as mean values over all subjects and source signals. Each panel
shows results for one sound property. The avarage ratings are plot-
ted for all synthesis conditions against rooms. Error bars indicate
inter-subject standard errors.
For naturalness (left panel) ratings differ strongly between
rooms. Whereas the BRIRs of the laboratory were rated to sound
most natural, lowest naturalness was perceived for the empty cham-
ber. BRIRs of the aula and lecture room were rated to have a
medium to high naturalness. Between synthesis conditions, almost
no differences are visible, and for most rooms, differences between
synthesized and measured BRIRs are very low. Moreover, for some
conditions even the synthesized BRIRs were rated to sound slightly
more natural than the measured one. This shows, that the proposed
simulation method is able to synthesize BRIRs that sound as natural
as measured ones. Remaining differences in perceived naturalness
between rooms might be due to familiarities of subjects with these
acoustic environments in daily life, since the Laboratory sounds as
dry as an ordinary living room, whereas the empty chamber sounds
rather unusual, even by the measured BRIR. This might also be
due to the unusual relation of its very small room size and its high
reverberation time (see Tab. 1).
For room size (right panel), again clear differences between
rooms are perceived. The order is well in accordance with reverber-
ation times and, except for the empty chamber, with the actual room
sizes (see Tab. 1). Differences within synthesis conditions and be-
tween synthesises and measurements are practically not existent.
Aula Empty ch. Lect. r. Lab.
high
2
3
4
5
6
low
Room
Naturalness
Aula Empty ch. Lect. r. Lab.
small
2
3
4
5
6
large
Room
Room size
N = 1
N = 3
ISM
meas.
Figure 3: Subjective sound property ratings of measured and synthe-
sized BRIRs for four rooms, averaged over all subjects and source
signals (speech and music). Error bars indicate inter-subject stan-
dard errors.)
This shows that firstly the simulation method is able to represent
different room sizes and secondly to achieve this independently
from maximum image source order as far as tested.
As a consequence of the experimental design, no direct map-
ping of synthesized BRIRs to actual rooms was performed. Given
that no head tracking was employed, a potential effect of head rota-
tions on the subjective ratings could so far not be assesed. Future
research will apply the system in a real-time environment and will
address this issue.
3.2. Evaluation of spatial properties of the extended FDN
3.2.1. Room- and synthesis conditions
Two rooms each with different configurations of source- and re-
ceiver positions as well as wall properties, were used to evaluate
the spatial reverb rendering. Fig. 1 depicts schematically the condi-
tions for one room. Tab. 2 specifies the room dimensions and the
absorption coefficients for (250, 500, 1k, 2k, 4k) Hz. While room 1
has an almost square base area, room 2 represents a long corridor.
Side wall absorption coefficients were specified in two different
ways: In the “closed” condition, all side wall absorption coeffi-
cients were equal as given in line 3 in Tab. 2. By this way, spatial
sound properties in azimuth can be investigated in dependence of
room geometry in connection to the positions of the receiver and
all virtual “reverb” sources. In the “open” condition, the left side
was completely open, meaning that no wall was existent. This was
technically represented by a broadband absorption coefficient of
0.99, whereas the absorption of all other side walls did not differ
from those of the closed condition. By this, the spatial rendering
was to be evaluated in a challenging condition for the model.
Also in both rooms and for the closed- and open condition,
the distance
d
of source and receiver to the left wall was chosen
to be 0.3 m or 5 m (see Fig. 1). The source was always in the
front direction of the receiver, yielding a direct sound with no
interaural differences. All differences are thus due to reflections
and reverberation.
Since no suitable real rooms were found for BRIR measure-
ments, the ISM with reflection order 20 was used as reference.
For all room conditions, the cube- and box condition were com-
pared against each other and the reference. ILDs and IACCs were
determined from the BRIR up to the time
min{T60(f)}
, where
T60(f)
is the frequency dependent reverberation time. Positive
ILDs indicate higher signal energy on the right.
Table 2: Specification of two virtual rooms used for evaluation of
the binauralization steps in the extended FDN. See text for further
explanation.
Dimensions room 1 (10.9, 10.8, 3.15) m
Dimensions room 2 (3.9, 30.0, 3.15) m
Absorption side walls (0.05, 0.10, 0.13, 0.16, 0,22)
Absorption open side 0.99
Absorption floor (0.03, 0.03, 0.03, 0.03, 0.02)
Absorption ceiling (0.70, 0.60, 0.70, 0.70, 0.50)
89
Proc. of the EAA Joint Symposium on Auralization and Ambisonics, Berlin, Germany, 3-5 April 2014
Table 3: Results of the evaluation of the FDN binauralization
steps: Comparison of cube and box method with purely ISM-created
BRIRs in terms of ILDs and IACCs. See text for explanation of
conditions.
ISM cube box
ILD (dB)
R 1, closed
d= 0.3m−0.3−0.71.3
d= 5.0m0.30.70.8
R 1, open
d= 0.3m0.90.91.8
d= 5.0m1.2 1.11.2
R 2, closed
d= 0.3m−0.6−0.23.1
d= 5.0m−0.60.73.7
R 2, open
d= 0.3m0.51.03.2
d= 5.0m0.51.03.0
ISM cube box
IACC
0.5 0.5 0.5
0.7 0.7 0.7
0.9 0.8 0.8
0.8 0.8 0.8
0.60.50.4
0.70.60.5
0.90.80.6
0.90.80.6
3.2.2. Results
The results for all room- and synthesis conditions are shown in
Tab. 3. Comparing results for the cube and box condition, values
that are closer to those of the ISM reference, with a difference of
at least 0.1 dB (ILD) or 0.1 (IACC) to the less matching condition,
are printed in bold face for clarity. To interpret the results just-
noticeable differences (JNDs) of ILD have to be considered as, e. g.
determined in [
17
] for musical instruments in several reverberant
conditions: 1.0–1.4 dB for
T60 = 1.3
s; 0.8–1.2 dB for
T60 = 0.8
s;
0.4–0.8 dB for the anechoic condition.
For room 1 (“R 1”) (closed) overall small absolute ILD values
are observed in the range of the JNDs. It has to be kept in mind,
that these ILDs originate from reflections only, given that the di-
rect sound was always located in the front direction. For the box
condition ILDs are larger and differ considerably more from the
reference. Clear mismatches to the reference ILD are obtained in
the close-to-wall position (d= 0.3m). Overall small or vanishing
ILDs are plausible for the closed conditions since all side walls
are equal in absorption coefficient. For the closed room 2 (“R 2”),
both ISM and cube condition show again small absolute ILD values.
However, ILDs for the box condition differ clearly from those of the
cube and reference conditions. This is not surprising because the
majority of virtual reverb sources lie clearly to the right hand side
of the receiver (see also scheme in Fig. 1). For the open versions of
the rooms, ILDs obtained from ISM-created BRIRs and the cube
condition have very similar values. Largest differences are again
obtained for the box condition in room 2.
The IACC results reveal overall no distinct differences between
the synthesis conditions for room 1. For room 2, where maximum
differences are 0.3, the cube condition yields IACCs that are closer
to those created by the ISM.
In conclusion, it can be said that the cube condition mostly cre-
ates spatially more realistic BRIRs in terms of ILD and IACC than
the box condition, especially when the room geometry and receiver
position are challenging. In addition, also an informal subjective
listening test yielded highest perceptive similarity between the cube
and the reference condition.
3.3. Evaluation of simulation of coupled rooms
3.3.1. Room- and synthesis conditions
The two adjacent rooms, an office and a corridor, acoustically cou-
pled by an open door, are specified in Tab. 4 in terms of dimensions
and absorption coefficients for (250, 500, 1k, 2k, 4k) Hz. The ar-
rangement of both rooms and positions of source S and receiver R
are depicted in Fig. 2. Two source positions were investigated. In
the “visible” condition the source is placed at the left end of the
double arrow, and in the “invisible” condition it is placed at the
right end. Measured BRIRs for two real rooms from which data
in Tab. 4 were obtained served as reference. The ISM condition
and the proposed hybrid method with
N= 3
and
N= 1
were
evaluated. In both hybrid conditions, the “cube” synthesis was
used.
Besides a comparison of the BRIRs in the time domain, ILDs
and IACCs were determined as described in sec. 3.2.1 and com-
pared with the reference.
Table 4: Specification of rooms used in the evaluation of the simu-
lation of coupled rooms.
Room 1 (corridor)
Dimensions (30.0, 1.94, 2.50) m
Absorption coeff. (0.16, 0.16, 0.13, 0.15, 0.17)
Room 2 (office)
Dimensions (4.43, 4.50, 3.00) m
Absorption coeff. (0.25, 0.30, 0.35, 0.32, 0.28)
3.3.2. Results
Fig. 4 shows normalized BRIR time signals for the measured (upper
panels) and synthesized (
N= 3
, lower panels) case on an arbitrar-
ily scaled ordinate. As expected for coupled rooms, the measured
BRIR in the invisible condition (rhs) shows a rising amplitude in
the beginning. This effect can hardly be observed in the simulated
BRIR. In this simple approach here, only one convolution of two
single RIRs was used which cannot mimick real coupling of the
rooms.
Tab. 5 shows ILDs and IACCs obtained from BRIRs of all
conditions. For all of them a clear dominance of sound energy
on the left is obtained (negative ILDs), which is primarily due
to the direction of the (virtual) direct sound (source S
0
in Fig. 2).
The ISM-created BRIRs, which can be assumed to simulate the
real rooms best, have indeed ILDs that are closest to those of the
measured BRIRs. The ILDs of the hybrid method BRIRs differ
maximally 3.4 dB from measurement condition, which is clearly
above the JND in reverberant conditions, at least for frontal source
positions [17].
For the IACC, all room- and synthesis conditions yield very
small values. A slightly higher accordance with the measurement
is obtained for the ISM synthesis, but it is questionable, whether
these differences were audible.
Concluding, the evaluation showed that this simple approach
has limitations if the acoustics of coupled rooms should be simu-
lated in a convincing way. Improvements should consider removal
of the direct path between the virtual source and the receiver in the
invisible condition. In a second step diffraction of the direct sound
90
Proc. of the EAA Joint Symposium on Auralization and Ambisonics, Berlin, Germany, 3-5 April 2014
amplitude
L
R
0 10 20 30 40 50
amplitude
time (ms)
0 10 20 30 40 50
time (ms)
Figure 4: BRIR time signals on arbitrary amplitude scale. Lhs:
visible condition, rhs: invisible condition. Upper panels: measured
BRIRs, lower panels: synthesis with N= 3.
Table 5: Results of the evaluation of the coupled-rooms simula-
tion. Comparison of ILDs and IACCs. See text for explanation of
conditions.
meas. ISM N= 3 N= 1
ILD (dB) visible −6.0−6.6−9.4−8.3
invisible −4.5−6.3−7.8−7.8
IACC visible 0.1 0.1 0.2 0.2
invisible 0.1 0.1 0.2 0.2
can be taken into account by inclusion of lowpass-filtered versions
of the direct path, depending on the geometric relation of the door
opening and the source and receiver positions.
4. SUMMARY AND CONCLUSIONS
A hybrid approach for synthesizing binaural room impulse re-
sponses of shoebox-shaped rooms was presented. It computes
geometrically exact early reflections using the image source model
up to a low reflection order, and approximates the reverberant tail
by a high efficient feedback delay network. The FDN was ex-
tended to enable a spatial reverb rendering, taking into account
room geometry as well as wall absorption and source- and receiver
positions.
The proposed simulation method was evaluated with respect to
different properties using subjective and technical measures. In a
subjective evaluation subjects rated the naturalness and room size
of measured and respectively synthesized BRIRs. The ratings show
that the simulation method is able to represent perceived naturalness
and room size very well and independently from maximum image
source order, whereas differences in these properties between rooms
are clearly conveyed.
For the extended FDN, two spatial reverb rendering techniques
(sec. 2.2) were compared in a technical evaluation assessing inter-
aural level differences and interaural cross correlation coefficients.
It was shown that synthesized spatial reverberation has better accor-
dance with purely ISM-created reference BRIRs if the reverberation
emitting virtual sound sources are equally distributed around the
listener’s head. In comparison, positioning these sources on the
actual room wall surfaces yielded worse results (sec. 3.2.2).
A first, simple approach to simulate the acoustics of two ad-
jacent coupled rooms was evaluated by comparing time signal
representations, ILDs and IACCs for measured and synthesized
BRIRs. While the results for this approach were not fully con-
vincing future improvement with refined approximations can be
expected.
In conclusion, the evaluation showed that the suggested compu-
tationally efficient approach for synthesizing binaural room impulse
responses is suited for applications where perceptual plausibility
and authenticity is acceptable.
5. ACKNOWLEDGEMENTS
This work was supported by the DFG FOR 1732 and the Cluster of
Excellence EXC 1077/1 “Hearing4all”.
6. REFERENCES
[1]
J. B. Allen and D. A. Berkley, “Image method for efficiently
simulating small-room acoustics,” J. Acoust. Soc. Am., vol.
66, no. 4, pp. 943–950, 1979.
[2]
B.-I. Dalenbäck, “Engineering principles and techniques in
room acoustics prediction,” in BNAM, Bergen, Norway, May
2010, 2010.
[3]
G. M. Naylor and J. H. Rindel, “Predicting Room Acoustical
Behaviour with the ODEON Computer Model,” in 124th ASA
meeting New Orleans, November 1992, 1992.
[4]
B.-I. Dalenbäck and M. Strömberg, “Real Time Walkthrough
Auralization – The First Year,” Tech. Rep., CATT (Dalen-
bäck), Valeo Graphics (Strömberg), 2010.
[5]
D. Schröder, F. Wefers, S. Pelzer, D. S. Rausch, M. Vorländer,
and T. Kuhlen, “Virtual Reality System at RWTH Aachen
University,” in Proceedings ICA 2010, 20th International
Congress on Acoustics: 23–27 August 2010, Sydney, New
South Wales, Australia, 2010.
[6]
A. Silzle, P. Novo, and H. Strauss, “IKA-SIM: A system to
generate auditory virtual environments,” in Audio Engineering
Society Convention 116, 2004.
[7]
A. Krokstad, S Strøm, and S. Sørsdal, “Calculating the acous-
tical room impulse response by the use of a ray tracing tech-
nique,” J. Sound Vib., vol. 8, no. 1, pp. 118–125, 1968.
[8]
Steven M. Schimmel, Martin F. Müller, and Norbert Dillier,
“A fast and accurate »shoebox« room acoustics simulator,”
Tech. Rep., 2009.
[9]
M. R. Schroeder, “Natural Sounding Artificial Reverberation,”
Journal of the Audio Engineering Society, vol. 10, no. 3, pp.
219–223, 1962.
[10]
J. Stautner and M. Puckette, “Designing Multi-Channel Rever-
berators,” Computer Music Journal, vol. 6, no. 1, pp. 52–65,
1982.
[11]
J.-M. Jot and A. Chaigne, “Digital delay networks for de-
signing artificial reverberators,” in 90th AES Convention,
1991.
[12]
T. Wendt, S. van de Par, and S. D. Ewert, “A computational
efficient and perceptually plausible algorithm for binaural
room impulse response simulation,” subm. to Journal of the
Audio Engineering Society.
91
Proc. of the EAA Joint Symposium on Auralization and Ambisonics, Berlin, Germany, 3-5 April 2014
[13]
J. Borish, “Extension of the image model to arbitrary poly-
hedra,” J. Acoust. Soc. Am., vol. 75, no. 6, pp. 1827–1836,
1984.
[14]
M. Jeub, M. Schäfer, and P. Vary, “A binaural room im-
pulse response database for the evaluation of dereverberation
algorithms,” Tech. Rep., 2009.
[15]
A. Farina, “Simultaneous measurement of impulse response
and distortion with a swept-sine technique,” in Audio Engi-
neering Society Convention 108, 2 2000.
[16]
G. Geißler and S. van de Par, “Messung von HRTF am Kun-
stkopf MK 2 von Cortex,” AG Akustik, Carl-von-Ossietzky-
Universität Oldenburg, 2012.
[17]
S. Klockgether and S. van de Par, “Just Noticable Differences
of Spatial Perception in Directly Manipulated Binaural Room
Impulse Responses,” in AIA/DAGA 2013, Merano, Italy,
2013.
92
